Statistical downscaling of maximum temperature under CMIP6 global climate models and evaluation of heat wave events using deep learning methods for Indo-Gangetic Plain

The rising global temperature is one of the primary concerns of the world as it impacts the economy, environment and healthcare of any country which are more pronounced a regional level. Assessment of regional impacts of climate change at a local level requires fine resolution of climate data for which a robust and fast downscaling method is needed. In this study, we use three deep learning-based methods, namely long short-term memory network (LSTM), deep neural network (DNN) and recurrent neural network (RNN), to downscale CMIP6 13 GCMS models data (1.25 degrees x 1.25 degrees resolution) global climate model (GCM) maximum temperature (Tmax) at a regional scale of 0.5 degrees x 0.5 degrees spatial resolution for the period 1991-2010 over the Indo-Gangetic Plain (IGP). In addition to the temperature prediction, heat wave events have been also analysed in the study. The study found that LSTM method performs better than DNN and RNN in downscaling of all GCM model datasets when evaluated against observed maximum temperature data from the India Meteorological Department (IMD) in terms of RMSE (0.9-3.5), average of all grid MAE value between (1.2 and 2.68), correlation (0.68-0.9) along with and spatiotemporal variability. LSTM also performed better in heat wave prediction over the region with similar temporal range (12-36 events) and spatial occurrence as compared to the observation (12-28 events). Overall, the study concludes that LSTM performs better than two methods for Indo-Gangetic Plain with best hyper parameter tuning. Hence, we propose to utilize a deep learning framework based on LSTM for downscaling GCM dataset at a finer resolution.